Single Atoms, Speedy Crows, Overshot Targets · Morning Gist
Monday, October 265 stories · 15 min
Single Atoms, Speedy Crows, Overshot Targets
A glass bead registers single atoms hitting it, crows out-decide primates, an ancient virus shapes chimp brains, and the UN plans for overshooting 1.5°C.
Picked from 56 new articles across 10 sources · arrow keys for other days
Today in 30 seconds
30s left
Yale physicists used a laser-levitated glass bead to detect kicks from individual gas atoms, each lasting a ten-trillionth of a second. story 1 ↓
Two carrion crows solved a moving-dots decision task in 235 milliseconds, twice as fast as primates and about as accurately. story 2 ↓
Removing a 5-to-7-million-year-old viral insertion from chimp stem cells switched on a gene humans use in brain development. story 3 ↓
About half the papers both neuroscience and AI cite most are software tools. AI's newest neuroscience classic dates from 2003. story 4 ↓
UNEP's best case now has warming peak at 1.8°C, then removing 320 billion tonnes of CO2 to return under 1.5°C by 2100. story 5 ↓
The FDA is weighing Kyverna's CAR T therapy, which could become the first treatment for stiff person syndrome.
1/5Science News· Emily Conover· Oct 22
Laser-levitated glass bead detects the push of single gas atoms
Yale physicists held a 100-nanometer glass sphere in a laser beam inside a vacuum and watched it twitch each time one gas atom or molecule hit it. The method could become a gauge for extremely low pressures.
Why it matters · Gas pressure is normally a statistical average over countless collisions. Here the individual collisions behind that average are counted one by one, a direct look at the kinetic theory of gases.
Know this first
Context from the editor, not the article · 20 sec
1 · Pressure is many collisions
Kinetic theory says gas pressure is the combined push of huge numbers of atoms or molecules bouncing off a surface. At everyday pressures the hits are so frequent that they merge into a smooth, steady force.
2 · Holding objects with light
A tightly focused laser beam can trap a tiny transparent particle near its brightest point, a technique called optical trapping, or optical tweezers. In a vacuum the particle floats with almost nothing touching it, so even very small disturbances show up.
Inside the vacuum chamber: the laser-trapped nanosphere is the bright dot floating between two metal pillars.
A kick you can count
2/5The Transmitter· Sarah Thau· Oct 22
Crows make simple visual decisions twice as fast as primates
Researchers at the University of Tübingen trained two crows on a standard moving-dots test used with monkeys for decades. The crows answered correctly about as often as primates do, in half the time.
Why it matters · Birds and mammals evolved their brains separately and built them very differently. If the bird version computes faster, the mammalian layered cortex is not the only way to build quick decisions, and maybe not the best one.
Know this first
Context from the editor, not the article · 32 sec
1 · Two ways to build a brain
Mammals do most of their higher thinking in the neocortex, a sheet of six stacked cell layers. Birds have no neocortex. Their pallium is arranged in clusters of neurons called nuclei. Even so, corvids such as crows count, recognize shapes and use tools.
2 · The moving-dots test
In this classic perceptual task, some of the dots on a screen drift together in one direction and the rest move at random. The subject reports which way the group is going. It has been used to study decisions in monkeys for about three decades.
3 · Perceptual vs. economic decisions
3/5Science News· Jake Buehler· Oct 23
Ancient virus DNA in chimp genomes silences a gene humans use in brain development
Swedish researchers cut leftover viral DNA out of chimpanzee stem cells. A nearby gene that is active in humans but silent in chimps switched on, raising activity in genes tied to neuron growth.
Why it matters · It shows a concrete way human and chimp brains could differ without differences in their proteins: an old infection changing when a gene switches on.
Know this first
Context from the editor, not the article · 27 sec
1 · Same proteins, different brains
Human brains are about three times larger than chimp brains, yet most of the two species' protein-coding genes are very similar. So researchers look at gene regulation: the DNA that controls when and where genes switch on.
2 · Viruses that write into genomes
Retroviruses insert their DNA into host cells. If that happens in sperm or egg cells, the viral DNA is inherited. These leftovers, called endogenous retroviruses, make up at least 8% of both the human and chimp genomes.
3 · How CRISPR is used here
CRISPR is a gene-editing tool that cuts DNA at chosen spots. It can delete a stretch of DNA or disable a gene, so scientists can test what that DNA does.
4/5The Transmitter· Renato Duarte· Oct 26
Citation study finds the neuroscience–AI exchange is growing but narrow
Neuroscientist Renato Duarte traced citations across 458 million papers. AI cites the brain more than chance predicts, but mostly old classics. Neuroscience mainly borrows AI's tools. A few big labs carry much of the traffic.
Why it matters · The claim that brain science and AI feed each other gets a lot of money and hype. This puts numbers on how much ideas actually flow between them, and which way.
Know this first
Context from the editor, not the article · 27 sec
1 · What NeuroAI promises
NeuroAI is the push, backed by figures like Geoffrey Hinton and Demis Hassabis, for neuroscience and AI to feed each other. The brain would supply principles and architectures, and AI would supply tools and theory for studying neural computation.
2 · Citations as a trail of ideas
Bibliometrics treats each reference one paper makes to another as a trace of influence. Counting citations that cross between fields over time gives a rough, imperfect picture of which field is borrowing from which.
3 · Neural networks' brain roots
Artificial neural networks, rebranded "deep learning" in the mid-2000s, were first loosely modeled on neurons and synapses. Today's systems resemble the brain mostly in concept, not in how they actually work.
5/5Ars Technica Science·
John Timmer
· Oct 22
UN concedes 1.5°C warming limit will be breached, sets an overshoot plan
For the first time, the UN Environment Programme says warming past 1.5°C can't be avoided. Its new goal is to exceed it briefly and by little, which means steep emissions cuts plus large-scale carbon removal.
Why it matters · The headline goal of the Paris Agreement now officially becomes a target to return to rather than one to stay under. That shifts the burden onto carbon removal, a technology that barely exists yet.
Know this first
Context from the editor, not the article · 30 sec
1 · The 1.5°C goal
Under the 2015 Paris Agreement, nearly every country pledged to try to keep global warming below 1.5°C above pre-industrial (roughly late-1800s) temperatures, and well below 2°C. Countries set their own emissions targets and are meant to tighten them over time.
2 · How 'crossing 1.5' is judged
One hot year doesn't count. The UN uses temperatures averaged over a couple of decades, so the limit is officially crossed only once the long-term average passes it. Single years, especially during El Niño, can go over sooner.
Physicists at Yale University have measured the push of a single gas atom. They held a glass sphere about 100 1 wide in a laser beam inside a nearly empty 2 and watched it jiggle each time one atom or molecule struck it. Yu-Han Tseng and colleagues reported the result in the October 23 issue of 3, a peer-reviewed physics journal.
Why it's hard
Gas pressure, like the air holding up a balloon, is normally the combined push of countless particles hitting all at once. Each particle does give its own small kick, but it is extremely faint. The sphere weighed about a quadrillionth of a gram, yet it was still roughly 10 million times heavier than each gas particle hitting it.
How it worked
Holding the sphere with laser light meant nothing mechanical was touching it. The vacuum kept collisions rare enough to tell apart. The team then let in small amounts of one of three heavy gases:
krypton, a gas of single atoms
xenon, also single atoms
4, a gas of molecules
By tracking the laser light reflecting off the sphere, they picked up the sudden small movements caused by individual impacts.
The smallest kicks measured were 5 (force multiplied by duration) of about a billionth of a 6 lasting one ten-trillionth of a second. By my rough estimate, not the article's, that is about the momentum a room-temperature xenon atom carries when it bounces straight back. Adding more gas raised both the pressure and the collision rate, which ties the single impacts to ordinary pressure.
What it might be for
Tseng says the method "could enable a very good way to sense pressure," especially at very low pressures like those in space or in vacuum chambers, where there is little gas to measure. Each gas also interacted with the sphere in its own way, so the setup could tell them apart. Tseng suggests a future device could study the makeup of gases in space or on other planets.
Caveats
The applications are proposals, not demonstrations. The article doesn't give the method's pressure range, how it compares with existing vacuum gauges, or how the gases were told apart. Building a field-ready instrument would be a separate project.
The result
A glass sphere about 100 1 across, held up by a laser inside a nearly empty 2, jiggled each time a single gas atom or molecule struck it. Yu-Han Tseng and colleagues at Yale University reported this in the October 23 issue of 3, a peer-reviewed physics journal. Science News covered the work in a short article, so many of the experiment's details aren't described here.
Why one atom is hard to notice
Usually we only meet the force of a gas in bulk. The air in a balloon keeps it inflated because enormous numbers of molecules hit the inside of the rubber every instant, and their pushes add up to a steady pressure. Each molecule still delivers its own small kick, but at ordinary pressures the kicks blur together, and each one is tiny compared with anything we can easily weigh or move.
The team's sphere weighed about a quadrillionth of a gram (10⁻¹⁸ kilograms). That is minuscule, but it is still roughly 10 million times the mass of a single gas particle. An analogy for scale: picture a moving ship struck by a grain of sand. To see such small kicks you need three things: a target light enough to move noticeably, few enough collisions that each one stands out, and a way to measure motion very precisely.
How the experiment worked
The researchers trapped the sphere with a laser in a vacuum chamber. This handled the first two needs. Nothing mechanical touches the sphere, and the near-vacuum means that only rare stray particles hit it. They then added small, controlled amounts of one of three heavy gases:
krypton, a gas of single atoms
xenon, also a gas of single atoms, and heavier than krypton
4, a gas made of molecules (one sulfur atom bonded to six fluorine atoms)
Heavy particles carry more momentum at a given temperature, so each one hits harder. That is likely why these gases were chosen, though the article doesn't say so. To track the sphere, the team monitored the laser light reflecting off it. Shifts in that light revealed sudden small movements of the sphere, each one the signature of a single collision.
The numbers
The kicks measured as 5 (force multiplied by the time it acts) as small as about a billionth of a 6 applied for one ten-trillionth of a second. That works out to roughly 10⁻²² newton-seconds.
A rough check, my own estimate rather than a figure from the article: a room-temperature xenon atom has a momentum of around 5 × 10⁻²³ kilogram-meters per second, and one that bounces straight back transfers about twice that. That matches the reported scale. Spread over a sphere of 10⁻¹⁸ kilograms, a kick of that size changes the sphere's speed by only about a tenth of a millimeter per second. That shows how sensitive the motion readout has to be.
A second result links the single kicks to ordinary pressure. As the researchers let more gas into the chamber, the collision rate rose. Pressure, in other words, could be read off by counting individual impacts.
Possible uses
Tseng says this "could enable a very good way to sense pressure," especially at very low pressures like those in space or in laboratory vacuum chambers. Measuring near-empty spaces is hard because there is so little gas to produce a signal. A device that counts collisions directly would sidestep that problem.
The system could also tell the three gases apart, because each interacted with the nanoparticle differently. Tseng suggests a future device might help work out what particles are present in space, or even on other planets: "You can imagine putting it on another planet and [using] that to study the environment there."
What the article leaves open
The report is peer-reviewed, but the applications are a forecast, not a demonstration. The article doesn't give:
the range of pressures the method works over
how its precision compares with existing vacuum gauges
exactly how the gases were told apart (for example, by size or pattern of kicks)
what fraction of collisions were detected rather than lost in noise
Turning a laboratory optical trap into a portable instrument, especially one that could go to another planet, is a separate engineering problem that the article doesn't address.
Remember
Holding the sphere with light, inside a vacuum, removed nearly all contact and noise, so single collisions became visible.
The collision rate rose as gas was added, linking individual impacts directly to measured pressure.
Low-pressure gauges and gas identification in space are proposed uses, not yet demonstrated.
A perceptual decision is a quick judgment about what the senses are reporting, such as left or right. An economic decision weighs the value of the options. This study tested only the first kind.
Cover image of the crow experiment. The article doesn't describe what it shows.
Carrion crows make simple visual decisions twice as fast as primates, humans included, and about as accurately. The study comes from Andreas Nieder's lab at the University of Tübingen and was published in Science on 22 October.
Different brains, same problem
Birds and mammals split more than 300 million years ago. Mammals think largely with the 1, a six-layered sheet of tissue. Birds have a 2 instead, with neurons grouped into clusters rather than layers. Crows still count, recognise shapes and use tools. Giving a crow and a monkey the same task is a direct way to see whether the different design changes how the problem gets solved.
The test
Two crows, Iggy and Lemmy, learned a task that has been used with monkeys for about 30 years. A cloud of white dots moves on a screen. Some drift together left or right, and the rest move randomly. The bird pecks the side the coherent dots are moving toward and gets a mealworm if it's right.
Nieder is clear that this measures quick perceptual judgment, not deliberation. Value-weighing "economic" decisions were not tested.
The numbers
After thousands of trials, the crows answered correctly in an average of 235 milliseconds. That is about half the time primates took in earlier studies, at a comparable accuracy of 74 percent. Both birds performed alike, which UCLA's Aaron Blaisdell found reassuring for such a small sample. Two crows is still a very small study, though, and the primate figures come from past work rather than a side-by-side test.
The team also recorded hundreds of neurons in the 3, the bird counterpart of the primate 4. Just 70 milliseconds after the dots appeared, NCL activity already predicted which side the crow would choose. Penn biologist Marc Schmidt, who was not involved, noted that birds evolved "the same types of computations," without a neocortex, "and in this case, to do them even better than mammals."
Why so fast?
Nieder calls this "an entirely wide-open research question" and offers three hypotheses:
Size: small brains mean shorter wiring between neurons.
Density: corvids pack two to three times as many neurons into the same volume.
Temperature: birds run warmer, which may speed neuron firing, ion channel opening and chemical signalling across synapses.
Next, Nieder wants to try harder decision tasks and to measure how confident the birds are. Blaisdell's question is whether the speed belongs to birds in general or to crows in particular.
Carrion crows can make simple visual decisions twice as fast as primates, humans included, and about as accurately. The study, led by animal physiologist Andreas Nieder at the University of Tübingen, was published in Science on 22 October. Nieder says it shows crow brains are "in the fast lane."
Why compare crows and primates at all
Birds and mammals last shared a common ancestor more than 300 million years ago, and their brains have taken different routes since. In mammals, higher cognition happens largely in the 1, a sheet of tissue organised into six stacked layers. Birds have no neocortex. Their equivalent region, the pallium, is a 2: neurons sit in clusters, or nuclei, instead of layers.
Corvids, the crow family, are clever all the same. Earlier work showed they count using neurons tuned to particular numbers, recognise geometric shapes and use tools. That makes them a natural test of how far brain function depends on brain structure. Give a crow and a monkey the same problem, and any difference in how they solve it says something about the hardware.
Marc Schmidt, a biologist at the University of Pennsylvania who was not involved in the work, describes this as "the power of the comparative approach": the two animals handle the same information in different ways and at different speeds.
The task
The team used a perceptual decision-making task that has been run with nonhuman primates for about 30 years. It is often called the random-dot motion task. A cloud of white dots appears on a screen. A portion of them move together, either left or right, and the rest move randomly. The subject must work out which way the coherent group is heading.
Two crows, named Iggy and Lemmy in the lab, learned to report their answer by pecking the matching side of the screen. A correct peck earned a mealworm.
Nieder stresses what the task measures. It is a quick perceptual judgment, not "decision-making in the pondering sense." The team did not test economic decisions, where an animal weighs the values of different options.
The results
After thousands of trials, the crows pecked the correct side in an average of 235 milliseconds. The article says that is twice as fast as primates given the same task in earlier studies. The crows' average accuracy was 74 percent, comparable to the primates'. So the extra speed did not come at the cost of more mistakes.
The two birds performed consistently with each other. Aaron Blaisdell of UCLA, also not involved, called that encouraging "with a small-N study like this." The authors write that this consistency could reflect an intrinsic biological capacity rather than one bird's quirks. Still, two animals is a very small sample, and the comparison with primates relies on earlier studies, not on primates tested side by side.
What the neurons showed
The team also recorded from hundreds of neurons in the 3, a bird brain region considered analogous to the primate 4, which handles executive functions such as decision-making. Seventy milliseconds after the dots appeared, NCL activity already predicted which side the crow would peck. That is well under a third of the way to the average response.
Schmidt calls this "very interesting." Birds lack a neocortex, he notes, "yet they evolved functional abilities to do the same types of computations. In this case, to do them even better than mammals."
Why might bird brains be faster?
Nieder calls this "an entirely wide-open research question" and offers three hypotheses, none of them tested here:
Size. Bird brains are small, so signals travel down shorter neuronal "wires." Less distance means less travel time.
Density. Corvids pack two to three times as many neurons into a given volume of tissue as primates do, which could shorten pathways or add computing power.
Temperature. Birds keep a higher steady body temperature. Warmer tissue can speed up neuronal firing, the opening of 5, and the movement of neurotransmitters across the 6, the gap between neurons.
As a rough way to picture the first two ideas (an analogy, not a claim from the study): a small, crowded city can be quicker to cross than a sprawling one, even if the cars move at the same speed.
What stays open
Nieder wants to try more complex decision tasks with crows, and to find ways of measuring how confident the birds are in their choices. Blaisdell asks whether the speed is "something that's avian versus mammal, or is it something about crows specifically?" Answering that would mean running the same test on other birds. Nieder doubts the deepest question, how bird brains manage to be "so terribly fast," will be settled within his career.
Remember
Neurons in the crows' NCL predicted each choice 70 milliseconds after the dots appeared, well before the peck.
Only two crows were tested, and the primate comparison relies on earlier studies rather than side-by-side testing.
Small brains, dense packing of neurons and warmer bodies are proposed explanations for the speed, but none has been tested.
An infection in an African ape 5 to 7 million years ago may have helped make chimpanzee brains different from ours. Researchers led by neuroscientist Johan Jakobsson at Lund University in Sweden report that viral DNA left over from that infection, still present in chimpanzees and missing in humans, keeps a gene switched off that is active in human brain cells. The study is published in the peer-reviewed journal Science Advances.
The puzzle
Human brains are three times the size of chimp brains and far more capable, yet most of the two species' protein-making genes are very similar. So the team looked at 1 instead: the DNA that decides when and where a gene turns on.
Viruses are one source of such regulatory changes. 2 insert their own DNA into the cells they infect. When that happens in reproductive cells, the viral DNA is passed to offspring. These inherited remnants, called 3, can no longer make infectious virus, but they can copy themselves around the genome and change how nearby genes behave. They make up at least 8% of the human and chimp genomes and have been linked to major evolutionary events, including the origin of the mammalian placenta.
The experiment
The viral remnant the team studied sits at 158 places in the chimp genome. It is absent from humans, and nobody knows why. Working with chimpanzee 4, the researchers used 5, a precise gene-editing tool, to cut the viral DNA out. A nearby gene called 6, normally silent in chimps, then switched on and began producing RNA. That matches human brain cells, where LINC00662 is active. In the reverse test, they turned LINC00662 off in human cells, copying the chimp state.
Whenever LINC00662 was active, in either species' cells, other genes became more active too. These included genes involved in how neurons develop and mature and how they form new connections.
How far it goes
The study does not show what silencing this gene network actually did to chimp brains. Everything was done in cells in the lab, not in living animals. Jakobsson calls it "one way" an ancient insertion could leave a lasting regulatory difference, and he expects other genetic and environmental factors to be involved too.
Evolutionary geneticist Ravinder Kanda of Oxford Brookes University, who was not involved, says the team "proved that absolutely" for these neurons but adds, "How far that goes, I don't think we can quite say yet." He suspects many more relevant retroviruses remain undiscovered.
An infection in an African ape somewhere between 5 million and 7 million years ago may have helped human and chimpanzee brains go separate ways. A team led by neuroscientist Johan Jakobsson at Lund University in Sweden reports that viral DNA left behind by that infection, still present in modern chimpanzees and missing in humans, keeps a gene switched off in chimp cells. In human brain cells that gene is active, and it appears to boost a network of genes involved in neuron development. The work was published October 23 in the peer-reviewed journal Science Advances.
The puzzle: similar genes, very different brains
Human brains are about three times larger than chimpanzee brains and more cognitively powerful. Yet, as Jakobsson puts it, "most of their protein-making genes are very similar." If the proteins are nearly the same, much of the difference has to come from somewhere else. The obvious place to look is 1: the stretches of DNA that control when, where and how strongly a gene is turned on. Two species can share a gene and still use it very differently.
Most earlier work on human–chimp brain differences focused on single-letter changes in the DNA sequence. Evolutionary biologist Giorgia Modenini of the University of Padua, who was not involved, says this study is a reminder that genomes change in other ways too.
Where viruses come in
2, the group that includes HIV, copy themselves by inserting their own DNA into a host cell's genome and using the cell's machinery to make more virus. Sometimes that inserted DNA ends up in an egg or sperm cell. Then it is inherited like any other DNA and can persist for millions of years.
These inherited remnants are called 3. They have mutated too much to produce infectious virus, but they still carry sequences that can affect nearby genes, and many have copied themselves repeatedly across the genome. Together they account for at least 8% of the human and chimp genomes. They have been linked to major evolutionary events before. The best-known example is the mammalian placenta, which relies in part on a protein borrowed from an ancient retrovirus.
The experiment
The team focused on one specific viral remnant. It came from an infection in an ancestral African ape 5 to 7 million years ago, around the time the human and chimp lineages split. In chimpanzees it sits at 158 separate locations in the genome. Humans do not have it, and the article notes that nobody yet knows why.
To test what the remnant does, the researchers took chimpanzee 4 and used 5, a gene-editing tool that makes precise cuts at chosen spots, to remove the viral DNA. With it gone, a nearby gene called 6 switched on and began producing RNA the team could detect. (The "LINC" prefix marks it as a long non-coding RNA gene. Its product is an RNA molecule that does its job without being translated into protein.) Normally this gene is silent in chimps, so removing the virus made the chimp cells look like human brain cells in this respect, since LINC00662 is active in humans.
They then ran the reverse experiment. In human cells they used CRISPR to shut LINC00662 off, copying the chimp condition.
The two-way design matters. Removing the insertion in chimp cells shows that the viral DNA is what silences the gene. Turning the gene off in human cells lets the researchers ask what the gene does when it is active and when it is not, using each species' own cellular background.
What the gene seems to do
Whenever LINC00662 was active, in either chimp or human cells, other genes became more active too. These included genes involved in how neurons develop and mature and in forming new connections between neurons. The likely chain is: viral insertion present → LINC00662 silenced → a set of brain-development genes runs at lower levels. In humans, who lack the insertion, the gene is on and that network is boosted.
An analogy: think of the viral DNA as a jammed switch. The wiring (the gene) is identical in both species, but in chimps a piece of foreign hardware holds the switch in the "off" position, and every light downstream stays dimmer.
What is and isn't established
The cell experiments are direct: remove the insertion and the gene turns on. Ravinder Kanda, an evolutionary geneticist at Oxford Brookes University who was not involved, says the team convincingly shows these viral leftovers influence developmental differences between human and chimp neurons. "They've proved that absolutely," he says. "How far that goes, I don't think we can quite say yet."
Several gaps remain:
Effect on real brains. The work was done in cells in the lab. It is not yet clear how silencing this gene network actually shaped chimpanzee brain biology, or whether its activity in humans contributed to larger or more capable brains.
One factor among many. Jakobsson calls the finding "one way an ancient viral [DNA] insertion could create a lasting difference" and expects other genetic and environmental changes to be involved too.
The missing insertion in humans. Why the remnant is in chimps at 158 sites but absent in humans is unexplained. The article does not say whether the human lineage lost it or never acquired it.
Scale. Kanda suggests this may be a small part of a larger story: "I think there are many, many more retroviruses to be discovered, and at the moment we've only touched the tip of the iceberg."
The result does not explain why human brains are bigger. What it offers is a demonstrated, testable mechanism: an inherited piece of viral DNA changing the regulation of a gene that sits upstream of neuron-development genes. That gives researchers a way to look for brain differences between species that are hidden in the regulatory DNA, not in the proteins.
Remember
The viral remnant sits at 158 sites in the chimp genome and is absent in humans, for unknown reasons.
The experiment ran both ways: removing the virus switched the gene on in chimp cells, and disabling the gene in human cells mimicked chimps.
All the work was done in lab-grown cells. The effect on real brain size or cognition is untested.
How often each field cites the other compared with chance, 1995–2025, with detected turning points.1 / 3
Neuroscientist Renato Duarte, writing in The Transmitter, tested whether neuroscience and AI really feed each other. That idea, branded 1, now has its own institutes and manifestos. Duarte used 2, an open database of publications, to trace citations across about 458 million works from 1995 to 2025. The analysis is open on GitHub but appears as an essay, not a peer-reviewed paper.
Correcting for size
AI's literature has exploded, so raw citation counts favor it. Duarte uses an 3 instead: how often one field cites the other compared with what chance would predict, given each field's size. By that measure, AI has cited neuroscience more than neuroscience has cited AI in every year studied.
The trends:
AI citing neuroscience declined from 2006 to 2019, during the rise of deep learning. It then climbed steeply from 2020 to 2025, reaching about double its 1995 level.
Neuroscience citing AI has risen since 2006. It sped up from 2016 to 2019 as machine-learning tools took over neural data analysis.
What AI borrows: about 90% of the neuroscience it cites is cognitive neuroscience, and under 5% is cellular or molecular work.
Duarte flags two caveats. First, 2024–25 data is unstable while OpenAlex rebuilds its records. Second, part of the post-2020 jump may reflect neuroscience shrinking as a share of AI's booming literature, rather than more real borrowing.
What the bridges are made of
Duarte also ranked about 59,000 papers that both fields cite. He used a 4, which rewards balance: a paper cited 100 times by each side beats one cited 1,000 times by AI and twice by neuroscience.
Tools make up about half of the top bridges: the 5 tumor-segmentation challenge, the statistics package , SciPy, ROC analysis and .
Duarte's argument
Duarte argues that AI's recent leaps owe the brain little, and that inspiration alone won't keep neuroscience relevant. He thinks the field needs to offer usable results. One target is how the brain learns efficiently on about 20 watts. Another is its methods for probing black-box systems, which could help researchers understand how AI models work.
Citations are an imperfect proxy for influence. And the most striking signal, the post-2020 surge, falls exactly where the data is least settled.
The question
Over the past decade, 1 has gathered institutes, endowments, manifestos and meetings. NeuroAI is the idea that neuroscience and artificial intelligence should drive each other forward in a "virtuous circle." Renato Duarte, a computational neuroscientist writing in The Transmitter, sets aside whether the two fields should feed each other. He asks whether they actually do.
To find out, he used 2, an open database of scholarly publications, covering about 458 million works and the 2.9 billion citation links between them. He tracked how often each field cited the other from 1995 to 2025. His short answer: the bridge is real and carrying more traffic. But it is narrow, it is heavy on tools, and a handful of institutions hold it up.
This is an essay backed by an open analysis. Code, data and methods are on GitHub. The article doesn't say it has been peer reviewed, so read the results as one researcher's careful but unrefereed study.
Why counting is tricky
The obvious approach is to count the citations crossing each way. But AI's literature has exploded in recent years, and a bigger field gets cited more simply because there is more of it. Duarte instead uses an 3: how often one field cites the other, compared with what you'd expect if references were handed out at random in proportion to each field's size.
An illustrative example (these numbers are not from the article): suppose neuroscience papers make up 5% of everything AI researchers could cite. If AI papers actually point 10% of their references at neuroscience, they cite it about twice as often as chance predicts.
Which way influence runs
Corrected for field size, the result surprised even Duarte: neuroscience has influenced AI more than AI has influenced neuroscience in every year from 1995 to 2025. That is, AI papers cite neuroscience more often, relative to chance, than the reverse.
The trends tell a more detailed story:
Neuroscience citing AI fell until about 2006, then rose steadily. It sped up from 2016 to 2019, as machine-learning tools spread through the analysis of neural data.
AI citing neuroscience rose gradually until 2006 and then leveled off. It declined significantly from 2006 to 2019, during the rise of deep learning. From 2020 to 2025 it climbed at the steepest rate in the whole series, reaching about double its 1995 level. That timing matches the NeuroAI manifestos.
What AI borrows is lopsided. About 90% of the neuroscience that AI papers cite is cognitive neuroscience, and less than 5% is cellular or molecular work.
If you skip the field-size correction, the picture flips. As a share of everything each field cites, neuroscience's citations of AI have led since 2018. Duarte attributes that crossover to how much AI work now exists, and he argues the corrected view is "the more honest" one.
Two caveats Duarte flags
First, OpenAlex is being rebuilt and is still ingesting and relinking recent records, so counts for 2024 and 2025 are unstable. The post-2020 upturn starts before those years, so it isn't an artifact. Its size, though, can't yet be measured.
Second, some of the recent jump in the odds ratio may come from neuroscience shrinking as a share of a booming AI literature, rather than AI researchers reaching for neuroscience more often. In raw shares, AI's citations to neuroscience did rise between 2020 and 2023, but more modestly than the odds ratio suggests. Whether this is a lasting revival, he writes, "remains to be seen."
Finding the bridges
Next, Duarte looked for the specific papers, people and institutions that both fields cite heavily. He ranked about 59,000 papers that had at least five citations from each side, scoring each one by the 4 of its citation counts from the two fields. The harmonic mean punishes imbalance instead of averaging it away. A paper cited 1,000 times by AI and twice by neuroscience scores about 4. A paper cited 100 times by each scores 100.
Papers. About half of the strongest bridges are tools, not ideas:
the 5 brain-tumor segmentation challenge
6, an R package for statistical models with mixed effects
SciPy, a scientific computing library for Python
ROC analysis, a standard way to evaluate classifiers
7, a machine-learning library
The other half are a classical neuroscience canon:
Bi and Poo's 1998 work on 8, the rule that a synapse strengthens or weakens depending on the precise order in which two neurons fire
Izhikevich's 2003 spiking neuron models
studies of visual attention and its saliency-based models, object recognition, receptive fields and sparse coding
work on working memory and short-term plasticity
The newest of these is from 2003. AI cites more neuroscience than ever, Duarte notes, but no paper from the past two decades has become essential reading in both fields.
People. The top authors mix AI-leaning researchers (Yoshua Bengio, Geoffrey Hinton, Antonio Torralba) with computational neuroscientists (Karl Friston, Peter Dayan, Wulfram Gerstner, Henry Markram, Christof Koch, Aude Oliva). A place on this list depends on pull from both fields at once, not on fame within one.
Institutions. Here the concentration is sharpest. Alphabet (Google plus DeepMind) and Microsoft rank first, and they are also the most AI-leaning names in the figure. MIT, UC Berkeley, Harvard, University College London and Stanford follow. The top 100 institutions account for nearly 40% of all bridge weight, while the top 100 authors account for less than 7%. In Duarte's reading, the bridge rests on institutional resources more than on particular individuals.
What Duarte argues follows
Duarte's view is that inspiration alone will no longer keep neuroscience relevant to AI. AI's recent leaps owe the brain little. He argues neuroscience must offer results AI can use on its own problems.
One opening is efficiency. The brain runs on roughly 20 watts, learns from a handful of examples and generalizes to new situations. Energy and sample efficiency, robustness and out-of-distribution generalization are concrete engineering targets, he writes, and neuroscience could earn its place by explaining the mechanisms behind them. Another is interpretability. Neuroscientists' methods for probing black-box systems from the inside could help researchers trying to understand how AI models work. He also warns that concentration in a few institutions tends to breed narrow consensus.
Open questions
Citations are a rough proxy for influence. They miss ideas that spread without credit and tools used without a reference. Sorting papers into "neuroscience" and "AI" is itself a judgment call, described in the repository. And the most interesting signal, the post-2020 surge, falls exactly where the data is least settled. The next few years of cleaner OpenAlex data should show whether it is a real renaissance or partly a statistical side effect of AI's growth.
Remember
Corrected for field size, AI has cited neuroscience more than neuroscience has cited AI every year since 1995.
No paper newer than 2003 has become essential reading in both fields.
The top 100 institutions hold nearly 40% of bridge-building citations, while the top 100 authors hold under 7%.
Carbon dioxide stays in the air for centuries, so warming only falls if CO2 is actively taken out. Overshoot scenarios let temperatures go past a target for a while, then rely on removal to bring them back down.
Image: Ars Technica Science
The United Nations Environment Programme (1), the UN agency that tracks how national climate pledges compare with what's needed, has said for the first time that the world won't keep warming below 1.5°C. "Decades of insufficient climate action makes exceedance of 1.5° C unavoidable," its new report says. Its replacement goal is to keep the 2 small and short, then bring temperatures back under 1.5°C by 2100.
This is a scenario-modelling report, not peer-reviewed research. Its odds come from climate models that the article doesn't describe.
Why now
Depending on which baseline and temperature record you use, the world is either just below 1.5°C or already past it. A strong 3, a periodic Pacific warming that lifts global temperatures, will almost certainly push every record over. The UN officially counts multi-decade averages, so the limit hasn't technically been crossed. But temperatures aren't falling much after hot years, and Ars Technica's John Timmer argues it is already too late in practice.
Where pledges lead
Under the 2015 4, countries set their own targets and were meant to tighten them over time. Most didn't follow through, and the US has withdrawn twice. UNEP projects:
Current policies: about 2.8°C by 2100
Every Paris pledge met: only 0.2°C better
Plus all 5 pledges (emitting no more than is removed): 2.1°C
Emissions were roughly flat from 2024 to 2025. Fossil fuel emissions actually rose, offset by better land management.
The new best case
UNEP's scenario gives a 66 percent chance that warming peaks at 1.8°C or less, and even odds of getting back below 1.5°C by 2100. It needs two things:
6, which means pulling CO2 out of the air and storing it. The target is 320 7 (billion tonnes) by 2100, from a technology still in its infancy. That's a bit over five years of today's emissions, spread across decades. Removal matters because CO2 lingers for centuries, so stopping emissions halts warming but doesn't reverse it.
Steep cuts: greenhouse gas emissions down 70 percent by 2040. Coal for electricity would fall 75 percent by 2035 and be phased out by 2040. Oil would fall 30 percent and gas 35 percent by 2040.
The gap
Current pledges cut emissions only about 12 percent by 2035. That leaves 16 gigatonnes of CO2 more than this pathway allows, compared with a 5-gigatonne gap for simply holding warming to 2°C. Coal is falling too slowly. Oil and gas are projected to rise for another decade. Solar and wind are growing fast, but demand is rising faster, so their share of the grid lags.
The article doesn't say which removal methods UNEP assumes, or whether overshoot damage to ice sheets and permafrost can be undone.
The United Nations Environment Programme (1), the UN agency that tracks the gap between national climate pledges and what's needed, has said for the first time that the world won't keep warming below 1.5°C. "Decades of insufficient climate action makes exceedance of 1.5° C unavoidable," its latest report says. "Whether it will be temporary crucially depends on decisions taken today."
The new goal, as Ars Technica's John Timmer explains it, is damage limitation. Keep the 2 as small and as short as possible, then bring temperatures back under 1.5°C by 2100. UNEP's best case peaks at about 1.8°C. Getting back down from there requires emissions cuts much steeper than anything countries have promised, plus hundreds of billions of tonnes of carbon dioxide pulled out of the atmosphere.
This is a policy and modelling report, not a peer-reviewed discovery. Its probabilities come from climate scenarios, and the article doesn't describe the models behind them.
Why the UN is only saying this now
Whether we've already passed 1.5°C depends on the baseline you pick for "pre-industrial" and which global temperature record you use. On some combinations we're just under; on others we're already over. The current 3, a periodic warming of the tropical Pacific that raises global temperatures, looks exceptionally strong, so it will almost certainly push readings past 1.5°C on every record.
The UN counts warming by averaging temperatures over a few decades. A hot year or two doesn't officially breach the limit, so on paper it hasn't been crossed. Timmer notes that temperatures haven't fallen much after recent record years, so the long-term average is probably entering the post-1.5° period now. Technically there's still a pathway that stays below the limit. In practice, Timmer argues, it is already too late.
How the Paris approach fell short
The 4 didn't impose cuts. Each country set its own emissions target, those targets were meant to ratchet up as technology and policy improved, and a periodic "5" (a formal global progress review) was supposed to check the results. UNEP concedes that "collectively countries have still not aligned policies with [agreed to] targets." The United States has withdrawn twice under Trump. Timmer notes that the gap between targets and actual policy would look even worse if US figures were still counted.
UNEP's end-of-century projections:
Current policies: about 2.8°C.
All Paris pledges, including the most ambitious conditional ones: only 0.2°C better, about 2.6°C.
Adding every 6 pledge made outside the Paris framework: 2.1°C. ("Net zero" means emitting no more than is removed.)
Warming keeps climbing into the next century in all three cases.
Recent data offer little relief. Emissions were roughly flat from 2024 to 2025. There may have been a slight decline, but it's within the uncertainty of zero change. Fossil fuel emissions actually rose, offset by better land-use management. To meet any of the targets, emissions needed to be falling by now.
The overshoot scenario
UNEP argues that every extra fraction of a degree, and every extra year above 1.5°C, raises the risk of irreversible change in ice sheets, permafrost carbon, ocean circulation and coral reefs, with "devastating human impacts." It frames the task as staying well below 2°C, minimizing "the duration and magnitude of overshoot," and setting a "credible return path to 1.5° C by the end of the century."
Its scenario gives a 66 percent chance that warming peaks no higher than 1.8°C, and even odds of being back below 1.5°C by 2100. That second number is a coin flip, not a guarantee.
The scenario has two parts.
1. 7. This means technologies or land practices that take CO2 out of the air and store it. The scenario needs 320 8 (billion tonnes) removed by 2100. That's a large target for a field Timmer calls "still in its infancy." He also offers a sense of scale: it's a bit more than five years of today's emissions, spread over decades. By simple arithmetic, that puts current annual emissions at roughly 60 gigatonnes. Removal matters because CO2 lingers for centuries. Cutting emissions to zero stops warming from rising but doesn't reverse it. Only taking carbon out does that. The article doesn't say which removal methods UNEP assumes, such as direct air capture, forests or soil carbon.
2. Rapid cuts. Greenhouse gas emissions would need to fall 70 percent between now and 2040. That means:
cutting coal used for electricity by 75 percent by 2035 and phasing it out entirely by 2040
cutting all uses of oil by 30 percent and gas by 35 percent by 2040
The size of the gap
Current pledges would cut emissions by only about 12 percent by 2035. That leaves the world emitting about 16 gigatonnes of CO2 more in 2035 than the minimal-overshoot pathway allows. The gap is much smaller for a pathway that simply holds warming to 2°C by 2100, at about 5 gigatonnes. That shows how much more the 1.5°C return path asks for.
Sector by sector, nothing is on track:
Coal is declining, but not nearly fast enough over the next decade.
Oil and gas need to start falling now but are projected to keep rising for a decade or more.
Solar and wind are growing fast enough for a scenario with reasonable conservation and efficiency. But electricity demand is rising so quickly that renewables are falling behind as a share of the grid.
One wildcard falls outside the report. It was prepared before the conflict with Iran disrupted oil and gas supplies, and there's no data yet on whether shortages have pushed people toward conservation or electrification.
What's left open
Several questions go unanswered. Can carbon removal grow from a niche industry to hundreds of gigatonnes in a few decades, and at what cost? Is a temporary overshoot actually reversible for the systems UNEP lists? Ice sheet loss and permafrost thaw may not simply undo themselves when temperatures fall. And an overshoot framing could become a reason to delay cuts now and rely on future removal. UNEP's plan needs the opposite: both at once, at full speed.
Timmer's closing point is pointed. Limiting the damage will take major efforts from many countries, and they are the same countries that got the world here by not taking the original goals seriously.
Remember
Even if every net-zero pledge is met, warming reaches about 2.1°C by 2100 and keeps rising afterward.
UNEP's best case only gives even odds of getting back below 1.5°C by 2100.
The 1.5°C return path needs a 70% emissions cut by 2040. Current pledges deliver about 12% by 2035.
Classic neuroscience makes up the rest. Examples include Bi and Poo's 1998 study of 8 (synapses changing strength depending on when neurons fire) and Izhikevich's 2003 spiking-neuron models. The newest of these classics dates from 2003.
Institutions concentrate the traffic. Alphabet and Microsoft lead, followed by MIT, Berkeley, Harvard, University College London and Stanford. The top 100 institutions hold nearly 40% of bridge weight, while the top 100 authors hold under 7%.
